Chapter 8 Chinese Text Processing

In this chapter, we will turn to the topic of Chinese text processing. In particular, we will discuss one of the most important issues in Chinese language processing, i.e., word segmentation.

When we discuss English parts-of-speech tagging in Chapter 5, it is easy to do the word tokenization in English because the word boundaries in English are more clearly delimited by whitespaces. Chinese, however, does not have whitespaces between characters, which leads to a serious problem for word tokenization.

We will look at the issues of word tokenization and talk about the most-often used library, jiebaR, for Chinese word segmentation. Also, we will include several case studies on Chinese text processing. In later Chapter 10, we will introduce another segmenter developed by the CKIP Group at the Academia Sinica. The CKIP Tagger seems to be the state-of-art tagger for Taiwan Mandarin, i.e., with more additional functionalities.

8.1 Chinese Word Segmenter jiebaR

8.1.1 Start

First, if you haven’t installed the library jiebaR, you may need to install it manually:

This is the version used for this tutorial.

## [1] '0.11'

Now let us take a look at a quick example. Let us assume that in our corpus, we have collected only one text document, with only a short paragraph.

There are two important steps in Chinese word segmentation:

  • Initilzie a segment object using worker()
  • Tokenize the texts into words using the function segment() with the designated segment object created earlier
##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     "指民眾"  
##  [7] "黨"       "不"       "分區"     "被"       "提名"     "人"      
## [13] "蔡壁如"   "黃"       "瀞"       "瑩"       "在昨"     "6"       
## [19] "日"       "才"       "請辭"     "是"       "為領"     "年終獎金"
## [25] "台灣民眾" "黨"       "主席"     "台北"     "市長"     "柯文"    
## [31] "哲"       "7"        "日"       "受訪"     "時則"     "說"      
## [37] "都"       "是"       "按"       "流程"     "走"       "不要"    
## [43] "把"       "人家"     "想得"     "這麼"     "壞"
## [1] "jiebar"  "segment" "jieba"

To word-tokenize the document, text, you first initialize a segment object, i.e., seg1, using worker() and feed this segment to segment(jiebar = seg1)and tokenize text into words.

8.1.2 Parameters Setting

There are many different parameters you can specify when you initialize the segmenter worker(). You may get more detail via the documentation ?worker. Some of the important arguments include:

  • user = ...: This argument is to specify the path to a user-defined dictionary
  • stop_word = ...: This argument is to specify the path to a stopword list
  • symbol = FALSE: Whether to return symbols (the default is FALSE)
  • bylines = FALSE: Whether to return a list or not (crucial if you are using tidytext::unnest_tokens())

Exercise 8.1 In our earlier example, when we created the segment object named seg1, we did not specify any arguments for worker(). Can you tell what the default settings are for the parameters of worker()?

Please try to create worker() with different settings (e.g., symbols = T, bylines = T) and see how the tokenization results differ from each other.

8.1.3 User-defined dictionary

From the above example, it is clear to see that some of the words are not correctly identified by the current segmenter: for example, 民眾黨, 不分區, 黃瀞瑩, 柯文哲.

It is always recommended to include a user-defined dictionary when tokenizing your texts because different corpora may have their own unique vocabulary (i.e., domain-specific lexicon).

This can be done when you initialize the segment object using worker(..., user = ...).

##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     "指"      
##  [7] "民眾黨"   "不分區"   "被"       "提名"     "人"       "蔡壁如"  
## [13] "黃瀞瑩"   "在昨"     "6"        "日"       "才"       "請辭"    
## [19] "是"       "為領"     "年終獎金" "台灣"     "民眾黨"   "主席"    
## [25] "台北"     "市長"     "柯文哲"   "7"        "日"       "受訪"    
## [31] "時則"     "說"       "都"       "是"       "按"       "流程"    
## [37] "走"       "不要"     "把"       "人家"     "想得"     "這麼"    
## [43] "壞"

The format of the user-defined dictionary is a text file, with one word per line. Also, the default encoding of the dictionary is UTF-8.

Please note that in Windows, the default encoding of a Chinese txt file created by Notepad may not be UTF-8. (Usually, it is encoded in big-5.)

Creating a user-defined dictionary may take a lot of time. You may consult 搜狗詞庫, which includes many domain-specific dictionaries created by others.

However, it should be noted that the format of the dictionaries is .scel. You may need to convert the .scel to .txt before you use it in jiebaR.

To do the coversion automatically, please consult the library cidian.

Also, you need to do the traditional-simplified Chinese conversion as well. For this, you may consult the library ropencc in R.

8.1.4 Stopwords

When you initialize the segment, you can also specify a stopword list, i.e., words that you do not need to include in the later analyses.

For example, in text mining, functional words are usually less informative, thus often excluded in the process of preprocessing.

##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     "指"      
##  [7] "民眾黨"   "不分區"   "被"       "提名"     "人"       "蔡壁如"  
## [13] "黃瀞瑩"   "在昨"     "6"        "才"       "請辭"     "為領"    
## [19] "年終獎金" "台灣"     "民眾黨"   "主席"     "台北"     "市長"    
## [25] "柯文哲"   "7"        "受訪"     "時則"     "說"       "按"      
## [31] "流程"     "走"       "不要"     "把"       "人家"     "想得"    
## [37] "這麼"     "壞"
Exercise 8.2 How do we quickly check which words were removed in segment(text, seg2) as compared to the results of segment(text, seg3)?
## [1] "日" "是" "都"

8.1.5 POS Tagging

So far we haven’t seen the parts-of-speech tags provided by the word segmenter. If you need the POS tags of the words, you need to specify the argument type = "tag" when you initialize the worker().

##          n         ns          n          x          n          n          x 
##     "綠黨"   "桃園市"     "議員"   "王浩宇"     "爆料"       "指"   "民眾黨" 
##          x          p          v          n          x          x          x 
##   "不分區"       "被"     "提名"       "人"   "蔡壁如"   "黃瀞瑩"     "在昨" 
##          x          d          v          x          n          x          x 
##        "6"       "才"     "請辭"     "為領" "年終獎金"     "台灣"   "民眾黨" 
##          n         ns          n          x          x          v          x 
##     "主席"     "台北"     "市長"   "柯文哲"        "7"     "受訪"     "時則" 
##         zg          p          n          v         df          p          n 
##       "說"       "按"     "流程"       "走"     "不要"       "把"     "人家" 
##          x          r          a 
##     "想得"     "這麼"       "壞"

The returned object is a named character vector, i.e., the POS tags of the words are included in the names of the vectors.

Every POS tagger has its own predefined tagset. The following table lists the annotations of the POS tagsets used in jiebaR:

Exercise 8.3 How do we convert the named word vector with POS tags returned by segment(text, seg4) into a long string as shown below?
## [1] "character"
## [1] "綠黨/n 桃園市/ns 議員/n 王浩宇/x 爆料/n 指/n 民眾黨/x 不分區/x 被/p 提名/v 人/n 蔡壁如/x 黃瀞瑩/x 在昨/x 6/x 才/d 請辭/v 為領/x 年終獎金/n 台灣/x 民眾黨/x 主席/n 台北/ns 市長/n 柯文哲/x 7/x 受訪/v 時則/x 說/zg 按/p 流程/n 走/v 不要/df 把/p 人家/n 想得/x 這麼/r 壞/a"

8.1.7 Reminders

When we use segment() as a tokenization method in the unnest_tokens(), it is very important to specify bylines = TRUE in worker().

This setting would make sure that segment() takes a text-based vector as input and returns a list of word-based vectors of the same length as output.

NB: When bylines = FALSE, segment() returns a vector.

## [[1]]
##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     "指民眾"  
##  [7] "黨"       "不"       "分區"     "被"       "提名"     "人"      
## [13] "蔡壁如"   "黃"       "瀞"       "瑩"       "在昨"     "6"       
## [19] "日"       "才"       "請辭"     "是"       "為領"     "年終獎金"
## [25] "台灣民眾" "黨"       "主席"     "台北"     "市長"     "柯文"    
## [31] "哲"       "7"        "日"       "受訪"     "時則"     "說"      
## [37] "都"       "是"       "按"       "流程"     "走"       "不要"    
## [43] "把"       "人家"     "想得"     "這麼"     "壞"
##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     "指民眾"  
##  [7] "黨"       "不"       "分區"     "被"       "提名"     "人"      
## [13] "蔡壁如"   "黃"       "瀞"       "瑩"       "在昨"     "6"       
## [19] "日"       "才"       "請辭"     "是"       "為領"     "年終獎金"
## [25] "台灣民眾" "黨"       "主席"     "台北"     "市長"     "柯文"    
## [31] "哲"       "7"        "日"       "受訪"     "時則"     "說"      
## [37] "都"       "是"       "按"       "流程"     "走"       "不要"    
## [43] "把"       "人家"     "想得"     "這麼"     "壞"
## [1] "list"
## [1] "character"

8.2 Chinese Text Analytics Pipeline

In Chapter 5, we have talked about the pipeline for English texts processing, as shown below:


Figure 8.1: English Text Analytics Flowchart


For Chinese texts, the pipeline is similar.

In the following Chinese Text Analytics Flowchar (Figure 8.2), I have highlighted the steps that are crucial to Chinese processing.

  • It is not recommended to use quanteda::summary() and quanteda::kwic() directly on the Chinese corpus object because the word tokenization in quanteda is not ideal (cf. dashed arrows in Figure 8.2).
  • It is recommended to use self-defined word segmenter for analysis. For processing under tidy structure framework, use own segmenter in unnest_tokens(); for processing under quanteda framework, create the tokens object, which is defined in quanteda as well.

Figure 8.2: Chinese Text Analytics Flowchart


It is important to note that when we specify a self-defined unnest_tokens(…,token=…) function, this function should take a character vector (i.e., a text-based vector) and return a list of character vectors (i.e., word-based vectors) of the same length.

In other words, when initializing the Chinese word segmenter, we need to specify the argument worker(…, byline = TRUE).


8.2.1 Tidy Structure Framework

  • So based on our simple corpus example above, we first transform the character vector text into a corpus object—text_corpus. With this, we can apply quanteda::summary() and quanteda::kwic() with the corpus object.
Exercise 8.4 Do you know why there are no tokens of concordance lines from kwic(text_corpus, pattern = "柯文哲")?
  • We can also transform the corpus object into a text-based TIBBLE using tidy(). Also, we generate an unique index for each row using row_number().
  • For word segmentation, we initialize the jiebaR segment object using worker().
  • Finally, we use unnest_tokens() to tokenize the text-based TIBBLE text_corpus_tidy into word-based TIBBLE text_corpus_tidy_word. That is, texts included in the text column are tokenized into words, which are unnested into rows of the word column in the new TIBBLE.

8.2.2 Quanteda Framework

  • Under the quanteda framework, we can also create the tokens object of the corpus and do kwic() search.
  • Most of the functions that work with corpus object can also work with tokens object in quanteda.

8.3 Comparing Tokenization Methods

quanteda also provides its own default word tokenization for Chinese texts. However, its default tokenization method does not allow us to add our own dictionary to the segmentation process, which renders the results less reliable. We can compare the the two results.

  • we can use quanteda::tokens() to see how quanteda tokenizes Chinese texts. The function returns a tokens object.
  • we can also use our own tokenization function segment() and convert the list to a tokens object using as.tokens(). (This of course will give us the same tokenization result as we get in the earlier unnest_tokens() because we are using the same segmenter my_seg.)

Now let’s compare the two resulting tokens objects:

  • These are the tokens based on self-defined segmenter:
##  [1] "綠黨"     "桃園市"   "議員"     "王浩宇"   "爆料"     ","      
##  [7] "指"       "民眾黨"   "不分區"   "被"       "提名"     "人"      
## [13] "蔡壁如"   "、"       "黃瀞瑩"   ","       "在昨"     "("      
## [19] "6"        ")"       "日"       "才"       "請辭"     "是"      
## [25] "為領"     "年終獎金" "。"       "台灣"     "民眾黨"   "主席"    
## [31] "、"       "台北"     "市長"     "柯文哲"   "7"        "日"      
## [37] "受訪"     "時則"     "說"       ","       "都"       "是"      
## [43] "按"       "流程"     "走"       ","       "不要"     "把"      
## [49] "人家"     "想得"     "這麼"     "壞"       "。"
  • These are the tokens based on default quanteda tokenizer:
##  [1] "綠黨"     "桃園市"   "議員"     "王"       "浩"       "宇"      
##  [7] "爆"       "料"       ","       "指"       "民眾"     "黨"      
## [13] "不"       "分區"     "被"       "提名"     "人"       "蔡"      
## [19] "壁"       "如"       "、"       "黃"       "瀞"       "瑩"      
## [25] ","       "在"       "昨"       "("       "6"        ")"      
## [31] "日"       "才"       "請辭"     "是"       "為"       "領"      
## [37] "年終獎金" "。"       "台灣"     "民眾"     "黨主席"   "、"      
## [43] "台北市"   "長"       "柯"       "文"       "哲"       "7"       
## [49] "日"       "受"       "訪"       "時"       "則"       "說"      
## [55] ","       "都是"     "按"       "流程"     "走"       ","      
## [61] "不要"     "把"       "人家"     "想得"     "這麼"     "壞"      
## [67] "。"

Therefore, for linguistic analysis, I would suggest to define own Chinese word segmenter, which is tailored to specific tasks/corpora.

8.4 Data

In the following sections, we look at a few more case studies of Chinese text processing using the news articles collected from Apple News as our example corpus. The dataset is available in our course dropbox drive: demo_data/applenews10000.tar.gz.

(This dataset was collected by Meng-Chen Wu when he was working on his MA thesis project with me years ago. The demo data here was a random sample of the original Apple News Corpus.)

8.5 Loading Text Data

When we need to load text data from external files (e.g., txt, tar.gz files), there is a simple and powerful R package for loading texts: readtext. The main function in this package, readtext(), which takes a file or a directory name from disk or a URL, and returns a type of data.frame that can be used directly with the corpus() constructor function in quanteda, to create a quanteda corpus object. In other words, the output from readtext can be directly passed on to the processing in the tidy structure framework (i.e., tidytext::unnest_tokens()).

The function readtext() works on:

  • text (.txt) files;
  • comma-separated-value (.csv) files;
  • XML formatted data;
  • data from the Facebook API, in JSON format;
  • data from the Twitter API, in JSON format; and
  • generic JSON data.

The corpus constructor command corpus() works directly on:

  • a vector of character objects, for instance that you have already loaded into the workspace using other tools;
  • a data.frame containing a text column and any other document-level metadata
  • the output of readtext::readtext()

8.6 quanteda::tokens() vs. jiebaR::segment()

In Chapter 4, we’ve seen that after we create a corpus object, we can apply kwic() to get the concordance lines of a particular word. At that time, we emphasized that this worked because quanteda underlyingly tokenized the texts behind the scene.

We can do the same the with Chinese texts as well:

In Section 8.3, I have made it clear that quanteda does have its own tokenization method (i.e., tokens()) for Chinese texts. It uses the tokenizer, stringi::stri_split_boundaries, which utilizes a library called ICU (International Components for Unicode) and the library uses dictionaries for segmentation of texts in Chinese.

The biggest problem is that we cannot add our own dictionary when using the default tokenization tokens() (at least I don’t know how). In other words, when we apply kwic() to apple_corpus, quanteda tokenizes the Chinese texts using its default tokenizer and perform the keyword-in-context search.

Like we did in Section 8.3, we can compare the word segmentation results between quanteda defaults and jiebaR (with own dictionary) with our current news corpus.

  • First we tokenize all texts in apple_corpus using jiebaR::segment() and the segmenter initilized with user-defined dictionary.
  • Second, we convert the returned list from segment() into a tokens object using as.tokens().
  • Third, we use quanteda default tokens() to convert the corpus object into tokens object.

Now we can compare the two versions of word segmentation. Let’s take a look at the first document:

## [1] 168
## [1] 148
##   [1] "《"     "蘋果"   "體育"   "》"     "即日起" "進行"   "虛擬"   "賭盤"  
##   [9] "擂台"   ","     "每名"   "受邀"   "參賽者" "進行"   "勝負"   "預測"  
##  [17] ","     "每周"   "結算"   "在"     "周二"   "公布"   ","     "累積"  
##  [25] "勝率"   "前"     "3"      "高"     "參賽者" "可"     "繼續"   "參賽"  
##  [33] ","     "單周"   "勝率"   "最高者" ","     "將"     "加封"   "「"    
##  [41] "蘋果"   "波神"   "」"     "頭銜"   "。"     "註"     ":"      "賭盤"  
##  [49] "賠率"   "如有"   "變動"   ","     "以"     "台灣"   "運彩"   "為主"  
##  [57] "。"     "\n"     "資料"   "來源"   ":"     "NBA"    "官網"   "http"  
##  [65] ":"      "/"      "/"      "www"    "."      "nba"    "."      "com"   
##  [73] "\n"     "\n"     "金塊"   "("      "客"     ")"      " "      "103"   
##  [81] ":"     "92"     " "      "76"     "人"     "騎士"   "("      "主"    
##  [89] ")"      " "      "88"     ":"     "82"     " "      "快艇"   "活塞"  
##  [97] "("      "客"     ")"      " "      "92"     ":"     "75"     " "     
## [105] "公牛"   "勇士"   "("      "客"     ")"      " "      "108"    ":"    
## [113] "82"     " "      "灰熊"   "熱火"   "("      "客"     ")"      " "     
## [121] "103"    ":"     "82"     " "      "灰狼"   "籃網"   "("      "客"    
## [129] ")"      " "      "90"     ":"     "82"     " "      "公鹿"   "溜"    
## [137] "馬"     "("      "客"     ")"      " "      "111"    ":"     "100"   
## [145] " "      "馬刺"   "國王"   "("      "客"     ")"      " "      "112"   
## [153] ":"     "102"    " "      "爵士"   "小牛"   "("      "客"     ")"     
## [161] " "      "108"    ":"     "106"    " "      "拓荒者" "\n"     "\n"
##   [1] "《"                 "蘋果"               "體育"              
##   [4] "》"                 "即日起"             "進行"              
##   [7] "虛擬"               "賭"                 "盤"                
##  [10] "擂台"               ","                 "每名"              
##  [13] "受邀"               "參賽者"             "進行"              
##  [16] "勝負"               "預測"               ","                
##  [19] "每周"               "結算"               "在"                
##  [22] "周二"               "公布"               ","                
##  [25] "累積"               "勝率"               "前"                
##  [28] "3"                  "高"                 "參賽者"            
##  [31] "可"                 "繼續"               "參賽"              
##  [34] ","                 "單"                 "周"                
##  [37] "勝率"               "最高"               "者"                
##  [40] ","                 "將"                 "加封"              
##  [43] "「"                 "蘋果"               "波"                
##  [46] "神"                 "」"                 "頭銜"              
##  [49] "。"                 "註"                 ":"                 
##  [52] "賭"                 "盤"                 "賠"                
##  [55] "率"                 "如有"               "變動"              
##  [58] ","                 "以"                 "台灣"              
##  [61] "運"                 "彩"                 "為主"              
##  [64] "。"                 "資料"               "來源"              
##  [67] ":"                 "NBA"                "官"                
##  [70] "網"                 "http://www.nba.com" "金塊"              
##  [73] "("                  "客"                 ")"                 
##  [76] "103"                ":"                 "92"                
##  [79] "76"                 "人"                 "騎士"              
##  [82] "("                  "主"                 ")"                 
##  [85] "88"                 ":"                 "82"                
##  [88] "快艇"               "活塞"               "("                 
##  [91] "客"                 ")"                  "92"                
##  [94] ":"                 "75"                 "公牛"              
##  [97] "勇士"               "("                  "客"                
## [100] ")"                  "108"                ":"                
## [103] "82"                 "灰"                 "熊"                
## [106] "熱火"               "("                  "客"                
## [109] ")"                  "103"                ":"                
## [112] "82"                 "灰"                 "狼"                
## [115] "籃網"               "("                  "客"                
## [118] ")"                  "90"                 ":"                
## [121] "82"                 "公鹿"               "溜"                
## [124] "馬"                 "("                  "客"                
## [127] ")"                  "111"                ":"                
## [130] "100"                "馬"                 "刺"                
## [133] "國王"               "("                  "客"                
## [136] ")"                  "112"                ":"                
## [139] "102"                "爵士"               "小牛"              
## [142] "("                  "客"                 ")"                 
## [145] "108"                ":"                 "106"               
## [148] "拓荒者"

Therefore, to work with the Chinese texts, if you need to utilize more advanced text-analytic functions provided by quanteda, please perform the word tokenization on the texts using your own word segmenter first and convert the object into a tokens, which can then be properly passed on to other functions in quanteda (e.g., dfm). (In other words, for Chinese text analytics, probably corpus object is less practical; rather, creating a tokens object of your corpus might be more useful.)

In the later demonstrations, we will use our own defined segmenter for word segmentation/tokenization.

8.7 Case Study 1: Word Frequency and Wordcloud

We follow the same steps as illstrated in the above flowchart 8.2 and deal with the Chinese texts using the tidy structure framework:

  • Load the corpus data using readtext() and convert it into an corpus object
  • Create a text-based tidy structure DF apple_corpus_tidy (i.e., a tibble)
  • Intialize a word segmenter using worker()
  • Tokenize the text-based data frame into a word-based tidy data frame using unnest_tokens()

These tokenization results should be the same as our earlier apple_tokens:

Creating unique indices for your data is very important. In corpus linguistic analysis, we often need to trace back to the original context where the word, phrase or sentence comes from. With all these unique indices, we can easily keep track of the sources of all tokenized linguistic units. Also, if the metadata of the source documents are available, these unique indices would allow us to connect the tokenized linguistic units to the metadata information (e.g., genres, registers, author profiles)

With a word-based tidy DF, we can easily generate a word frequency list as well as a wordcloud to have a quick overview of the word distribution in the corpus.

8.8 Case Study 2: Patterns

In this case study, we are looking at a more complex example. In corpus linguistic analysis, we often need to extract a particular pattern from the texts. In order to retrieve the target patterns at a high accuracy rate, we often need to make use of the additional annotations provided by the corpus. The most often-used information is the parts-of-speech tags of words. So here we demonstrate how to add POS tags information to our current tidy corpus design.

8.8.1 Define Own Tokenization Functions

  • We define two tokenization functions:

    • chinese_chunk_tokenizer(): This function tokenizes a document text into a series of inter-punctuation units. We refer to these units as sentence-like chunks.
    • chinese_word_tokenizer(): This function tokenizes a text into a vector of “word/tag” tokens.
  • Initialize worker()

When initilizing the word segmenter worker(), remember to specify the argument type = "tag" to get POS tags. Also, we specify own dictionary (user = ...) and keep symbols (symbol=T) when doing the word tokenization.

We can try our self-defined functions with one text from the corpus:

## [1] "《蘋果體育》即日起進行虛擬賭盤擂台,每名受邀參賽者進行勝負預測,每周結算在周二公布,累積勝率前3高參賽者可繼續參賽,單周勝率最高者,將加封「蘋果波神」頭銜。註:賭盤賠率如有變動,以台灣運彩為主。\n資料來源:NBA官網http://www.nba.com\n\n金塊(客) 103:92 76人騎士(主) 88:82 快艇活塞(客) 92:75 公牛勇士(客) 108:82 灰熊熱火(客) 103:82 灰狼籃網(客) 90:82 公鹿溜馬(客) 111:100 馬刺國王(客) 112:102 爵士小牛(客) 108:106 拓荒者\n\n"
## [[1]]
##  [1] ""                           "蘋果體育"                  
##  [3] "即日起進行虛擬賭盤擂台"     "每名受邀參賽者進行勝負預測"
##  [5] "每周結算在周二公布"         "累積勝率前"                
##  [7] "高參賽者可繼續參賽"         "單周勝率最高者"            
##  [9] "將加封"                     "蘋果波神"                  
## [11] "頭銜"                       "註"                        
## [13] "賭盤賠率如有變動"           "以台灣運彩為主"            
## [15] "資料來源"                   "官網"                      
## [17] "金塊"                       "客"                        
## [19] "人騎士"                     "主"                        
## [21] "快艇活塞"                   "客"                        
## [23] "公牛勇士"                   "客"                        
## [25] "灰熊熱火"                   "客"                        
## [27] "灰狼籃網"                   "客"                        
## [29] "公鹿溜馬"                   "客"                        
## [31] "馬刺國王"                   "客"                        
## [33] "爵士小牛"                   "客"                        
## [35] "拓荒者"                     ""
## [[1]]
##   [1] "《/x"      "蘋果/n"    "體育/vn"   "》/x"      "即日起/l"  "進行/v"   
##   [7] "虛擬/v"    "賭盤/x"    "擂台/v"    ",/x"      "每名/x"    "受邀/v"   
##  [13] "參賽者/n"  "進行/v"    "勝負/v"    "預測/vn"   ",/x"      "每周/r"   
##  [19] "結算/v"    "在/p"      "周二/t"    "公布/v"    ",/x"      "累積/v"   
##  [25] "勝率/n"    "前/f"      "3/x"       "高/a"      "參賽者/n"  "可/v"     
##  [31] "繼續/v"    "參賽/n"    ",/x"      "單周/x"    "勝率/n"    "最高者/n" 
##  [37] ",/x"      "將/zg"     "加封/v"    "「/x"      "蘋果/n"    "波神/x"   
##  [43] "」/x"      "頭銜/n"    "。/x"      "註/x"      ":/x"       "賭盤/x"   
##  [49] "賠率/n"    "如有/c"    "變動/vn"   ",/x"      "以/p"      "台灣/x"   
##  [55] "運彩/x"    "為主/x"    "。/x"      "\n/x"      "資料/n"    "來源/n"   
##  [61] ":/x"      "NBA/eng"   "官網/x"    "http/eng"  ":/x"       "//x"      
##  [67] "//x"       "www/eng"   "./x"       "nba/eng"   "./x"       "com/eng"  
##  [73] "\n/x"      "\n/x"      "金塊/n"    "(/x"       "客/n"      ")/x"      
##  [79] " /x"       "103/m"     ":/x"      "92/m"      " /x"       "76/m"     
##  [85] "人/n"      "騎士/n"    "(/x"       "主/b"      ")/x"       " /x"      
##  [91] "88/m"      ":/x"      "82/m"      " /x"       "快艇/n"    "活塞/vn"  
##  [97] "(/x"       "客/n"      ")/x"       " /x"       "92/m"      ":/x"     
## [103] "75/m"      " /x"       "公牛/n"    "勇士/n"    "(/x"       "客/n"     
## [109] ")/x"       " /x"       "108/m"     ":/x"      "82/m"      " /x"      
## [115] "灰熊/x"    "熱火/n"    "(/x"       "客/n"      ")/x"       " /x"      
## [121] "103/m"     ":/x"      "82/m"      " /x"       "灰狼/n"    "籃網/n"   
## [127] "(/x"       "客/n"      ")/x"       " /x"       "90/m"      ":/x"     
## [133] "82/m"      " /x"       "公鹿/n"    "溜/v"      "馬/n"      "(/x"      
## [139] "客/n"      ")/x"       " /x"       "111/m"     ":/x"      "100/m"    
## [145] " /x"       "馬刺/nr"   "國王/n"    "(/x"       "客/n"      ")/x"      
## [151] " /x"       "112/m"     ":/x"      "102/m"     " /x"       "爵士/n"   
## [157] "小牛/n"    "(/x"       "客/n"      ")/x"       " /x"       "108/m"    
## [163] ":/x"      "106/m"     " /x"       "拓荒者/nr" "\n/x"      "\n/x"

In the above example, we adopt a very naive approach by treating any linguistic unit in-between the punctuation marks as a possible sentence-like unit. This can be controversial to many grammarians and syntaticians. However, in practice, this may not be a bad choice as it will become obvious when we extract patterns.

For more information related to the unicode range for the punctuations in CJK languages, please see this SO discussion thread.

8.8.2 Transform Text-Based to Token-Based Data Frame

Now we can apply our self-defined tokenization functions to the text-based DF apple_df.

We first unnest_tokens() the text-based DF into a chunk-based DF using the tokenizer chinese_chunk_tokenizer(). Then we transform the chunk-based DF into a word-based DF using chinese_word_tokenizer().

##    user  system elapsed 
##  17.087   0.261  17.380
## [1] 2375674       4

The word-based data frame now has parts-of-speech tags for every word in the corpus. Based on the word_id, chunk_id, and doc_id, we can easily keep track of their source documents as well.

Now based on the word-based data frame, we create a chunk-based data frame again by concatenating all word/tag in a chunk into a long string. In our earlier chunk tokenization, we only split texts into chunks without performing the word segmentation and POS tagging yet. The word boundary and POS information is only available when we perform the word tokenization using chinese_word_tokenizer. Therefore, to get a chunk with both words and POS tags, we can concatenate “word/tag” tokens into a long string on a chunk basis.

##    user  system elapsed 
##   4.696   0.189   4.889
## [1] 550504      3

This step may seem a bit redundant at the first sight. Why do we need to combine “word/tag” tokens into a longer string AGAIN??

It is indeed NOT necessary but the chunk-based data frame would be more useful for further construction/pattern analysis. Syntactic patterns/constructions often span word boundaries but stay within a sentence frame.

Exercise 8.5 In the above demonstration, we perform two rounds of tokenizations in the text preprocessing: we first tokenize the texts into chunks using unnest_tokens(..., token = chinese_chunk_tokenizer); we then tokenize the chunks into words using unnest_tokens(..., token = chinese_word_tokenizer).

However, this may be unnecessary. It is also possible to get the POS-tagged version of the entire text with words and tags available without chunking. This would probably be more efficient. How would you create one additional column like text_tag directly from the text-based DF apple_df, as shown below? (Please note that it is still a text-based DF and therefore probably you can use mutate).
##    user  system elapsed 
##   7.148   0.118   7.272

8.8.3 BEI Construction

This section will show you how we can make use of the chunk-based corpus data frame with POS tags. I would like to illustrate its usefulness with a case study: 被 + ... Construction.

After we tokenize the text-based tidy corpus into a inter-punctuation-unit-based (IPU), i.e., chunk-based data frame, we can make use of the words as well as their parts-of-speech tags to extract the target pattern we are interested: 被 + ... Construction.

The data retrieval process is now very straighforward: we only need to go through all the chunks in the corpus object and see if our target pattern matches any of these chunks. The assumption is that: the BEI-Construction will NOT span different chunks.

In the following example, we:

  • define a regular expression \\b被/p\\s([^/]+/[^\\s]+\\s)*?[^/]+/v for BEI-Construction, i.e., 被 + VERB
  • use unnest_tokens() and str_extract_all() to extract target patterns

Please check Chapter 5 Parts of Speech Tagging on evaluating the quality of the data retrieved by a regular expression (i.e., precision and recall).

To have a more in-depth analysis of BEI construction, we like to automatically extract the verb used in the BEI construction.


Exercise 8.6 When you take a closer look at the resulting word cloud above, you would see the copular verb 是 showing up in the graph, which is counter to our native speaker intuition. How do you check the instances of these 是 tokens? After you examine these cases, what do you think may be the source of the problem?
Exercise 8.7 To more properly evaluate the quality of the pattern queries, it would be great if we still have the original chunk texts available in the resulting data frame result_bei. How do we keep this information? That is, please have one column in result_bei, which shows the original chunk texts from which the construction token is extracted.

Exercise 8.8 Please use the apple_chunk_df as your tidy corpus and extract Chinese particle constructions of ... 外/內/中. Usually a space particle construction like these consists of a landmark NP (LM) and the space particle (SP). For example, in 任期內, 任期 is the landmark NP and is the space particle. In this exercise, we will naively assume that the word directly preceding the space particle is our landmark NP head noun. So please (a) extract all concordance lines with these space particles and (b) at the same time identify their respective SP and LM, as shown below.

Exercise 8.9 Following Exercise 8.8, please generate a frequency list of the LMs for each spac particle. Show us the top 10 LMs of each space particle and arrange the frequencies of the LMs in a descending order, as shown below.

Exercise 8.10 Following Exercise 8.9, for each space particle, please create a word cloud of its co-occuring LMs based on the top 100 LMs of each particle.

PS: The word frequencies in the word clouds shown below are on a log scale.


Exercise 8.11 Based on the word clouds provided in Exercise 8.9, do you find any bizarre cases? Can you tell us why? What would be the problems? Or what did we do wrong in the text preprocessing that may lead to these cases?

Please discuss these issues in relation to the steps in our data processing, i.e., word segmentation, POS tagging, and pattern retrievals.

8.9 Case Study 3: Lexical Bundles

8.9.1 N-grams Extraction

With word boundaries, we can also analyze the recurrent multiword units in Chinese news. Here let’s take a look at recurrent four-grams. As we discussed in Chapter 4, a multiword unit can be defined based on at least two metrics:

  • the frequency of the whole multiword unit (i.e., frequency)
  • the number of texts where the multiword unit is observed (i.e., dispersion)

As the default n-gram tokenization in unnest_tokens() only works with the English data, we start this task by defining our own token function ngram_chi() to extract Chinese n-grams.

This ngram_chi() takes ONE text (scalar) as an input, and returns a vector of n-grams. Most importantly, this function assumes that in the text string, each word token is delimited by a whitespace (i.e., a word-segmented text!!)

## [1] "這_是"     "是_一個"   "一個_測試" "測試_的"   "的_句子"
## [1] "這_是_一個_測試"   "是_一個_測試_的"   "一個_測試_的_句子"
## [1] "這 是 一個 測試 的"   "是 一個 測試 的 句子"
## [1] ""

We vectorize the function ngram_chi(). This step is important because in unnest_tokens() the self-defined token function should take a text-based vector as input and return a list of token-based vectors of the same length as the output (cf. Section 8.2).


Vectorized functions are a very useful feature of R, but programmers who are used to other languages often have trouble with this concept at first. A vectorized function works not just on a single value, but on a whole vector of values at the same time.

In our first defined ngram_chi function, it takes one text vector as an input and processes it one at a time. However, we would expect ngram_chi to process a vector of texts (i.e., multiple texts) at the same time and return a list of resulting ngrams vectors at the same time. Therefore, we use Vectorize() as a wrapper to vectorize our function and specifically tell R that the argument text is vectorized, i.e., process each value in the text argument in the same way.


Now we can tokenize our corpus into n-grams using our own token function vngram_chi() and the unnest_tokens(). In this case study, we demonstrate the analysis of four-grams in our Apple News corpus.

  • We first remove all POS tags in apple_chunk_df$chunk because n-grams do not need the POS tags
  • We then transform the chunk-based data frame apple_chunk_df into a n-gram-based data frame using unnest_tokens(...) with self-defined token function
  • We remove chunks with no target n-grams extracted (Chunks with less than four words will have NO four-grams extracted.)
##    user  system elapsed 
##  83.018   0.229  83.335
## [1] 974022      3

Exercise 8.12 Because n-grams extraction often requires no POS tags, it is not necessary (or redundant) to perform the POS tagging first and then remove the tags again indeed. For this task, we can split the raw text corpus into chunks and then do the word segmentation as well as n-grams extraction at the same time. It is also possible to create a self-defined function like chinese_ngram_tokenizer, which takes a simpler word segmenter, and directly get the apple_ngram from apple_df

Please define a function chinese_ngram_tokenizer to make the following tokenization codes work so that we can generate ngrams directly from apple_df. The following codes should produce the same result as the above apple_ngram.
##    user  system elapsed 
##  40.731   0.100  40.890
## [1] 974022
## [1] 974022

8.9.2 Frequency and Dispersion

Now that we have the four-grams-based DF, we can compute their token frequencies and document frequencies in the corpus using the normal data manipulation tricks.

We set cut-offs for four-grams: dispersion >= 5 (i.e., four-grams that occur in at least five different documents)

Please take a look at the four-grams, both arranged by frequency and dispersion:

We can also look at four-grams with particular lexical words:


Exercise 8.13 In the above example, if we are only interested in the four-grams with the word , how can we revise the regular expression so that we can get rid of tokens like ngrams with 以及, 以上 etc.

8.10 Afterwords

Tokenizations are complex in Chinese text processing. Many factors may need to be taken into account when determining the right tokenization method. While word segmentation is almost a necessary step in Chinese computational text analytics, several important questions may also be relevant to the data processing methods:

  1. Do you need the parts-of-speech tags of words in your research?
  2. What is the base linguistic unit you would like to work with? Texts? Chunks? Sentences? N-grams? Words?
  3. Do you need non-word tokens such as symbols, punctuations, or numbers in your analysis?

Your answers to the above questions should help you determine the most effective structure of the tokenization methods for your data.